✿ Anime / Sentiment Atlas

SAVED REVIEW ANALYSIS

Compare sentiment predictions
with anime review verdicts.

186 reviews across nine titles with collected reviews. One Piece supplies 100 reviews, so this snapshot does not represent anime audiences evenly.

A research project by Harsh Raj · MyAnimeList review verdicts

Loading research results…

Loading saved snapshot…

01 / THE VIEWERS

What did they really think?

These are reviewers’ own verdicts.
They are different from a model’s prediction.

Review verdicts

THE SENTIMENT GAP

Praise doesn’t always
mean a recommendation.

Why this happens ↗

Comments and reviews by anime

AnimeSite score / 10Forum commentsLabelled reviews

Click a title for its sample details. Forum comments are the expanded AniList and MyAnimeList dataset, capped at 5,000 per anime. Labelled reviews are a separate sample used to evaluate recommendation accuracy; their counts do not include forum comments. The site score reflects MyAnimeList’s wider audience.

FORUM / COMMUNITY VOICES

Beyond the review.

Up to 5,000 unique forum comments per anime.
Model predictions, separate from review verdicts.

Loading discussion snapshot…

AnimeCommentsPositiveNeutralNegative
How discussion sentiment works

AniList threads linked to each anime are sampled newest thread first. MyAnimeList anime-linked topics provide an additional source when API access is configured. Replies and nested replies are deduplicated by comment ID, up to 5,000 per title. VADER predicts positive, neutral and negative language using compound cut-offs of ±0.05. Comments can discuss characters, other viewers or unrelated subjects; these percentages do not measure recommendation rates or validated accuracy. The combined cap is 5,000 across both sources, with AniList collected first. Samples under 100 are flagged; zero counts mean no discussions found in the collected sources. Source access and completion are shown separately. Counts cannot establish that a title has no discussions elsewhere.

02 / THE MODELS

Same reviews. Different readings.

Binary evaluation uses Recommended vs Not Recommended.
Mixed Feelings is excluded from this comparison.

Read the evaluation limits

Models are evaluated with folds separated by anime, so the trained model does not see reviews from the held-out show. Tuned cut-offs use training folds only. This small snapshot is unevenly distributed across shows; differences between methods should be read alongside confidence intervals, not treated as a universal ranking.

View the full saved report ↗

03 / THE CONTEXT

Read the review.
Respect the verdict.

Re-run in Colab ↗
  1. 01

    Collect recent reviews

    The pipeline samples MyAnimeList’s top airing titles and keeps the reviewer’s Recommended, Mixed Feelings, or Not Recommended verdict. Preliminary reviews may be included.

  2. 02

    Compare words with conclusions

    TextBlob and VADER score the language. A TF-IDF model learns from verdicts. A negative review may still praise animation, music, or earlier episodes.

  3. 03

    Test on unseen shows

    Evaluation splits by anime and resamples whole shows for uncertainty. This frontend reads saved aggregate outputs; it does not run inference or collect user text.

SNAPSHOT / TITLE DETAILS

View on MyAnimeList ↗